Maximum-likelihood blind equalization of multiple FIR channels

We pursue our iterative quadratic maximum likelihood (IQML) approach to blind estimation of multiple FIR channels. We use a parameterization of the noise subspace in terms of linear prediction quantities. This parameterization is robust w.r.t. a channel length mismatch. Specifically, when the channel length is overestimated, no problems occur. Underestimation leads to a reduced-order channel estimate. We introduce two matched filter bounds (MFBs) to characterize the performance of receivers using reduced-order channel models. The first one (MFB1) uses the channel model to perform the spatio-temporal matched filtering that yields data reduction from multichannel to single-channel form. The rest of the processing remains optimal. MFB2 on the other hand bounds the performance of the Viterbi algorithm with the reduced channel model. It is shown that the reduced model provided by IQML is the one that maximizes MFB1. We also propose some low complexity techniques for obtaining consistent estimates with which to initialize IQML